A new arXiv preprint investigates the optimization landscape of the maximum mean discrepancy (MMD) when the optimizing measures are weighted atomic measures. This setting is relevant to algorithms that approximate a target distribution using a finite set of weighted points, and the paper focuses on how the landscape changes as the number of atoms grows.

The authors prove results for the one-dimensional Gaussian kernel, assuming the target distribution is compactly supported. In this case, they show that with sufficiently many atoms, the optimization landscape acquires desirable properties—though the abstract does not specify the exact nature of these properties beyond the general claim.

Because the source is only the abstract, the precise conditions and conclusions remain partially open. Still, the work adds to a growing body of theory on why overparameterised distributional optimization tends to avoid poor local minima.